Large language models unlock large text corpora in the search for data on medicinal plants and fungi

Societal Impact Statement Millions of people worldwide depend on plants and fungi for their healthcare needs, yet our understanding of which species are used medicinally—and how—remains fragmented and incomplete. As part of Kew's Plants for Health service, we developed a pipeline to mine large text corpora for information on natural products, initially focusing on the medicinal properties of plants and fungi. We show that large language models can be leveraged to efficiently mine this information from the CORE dataset of open access papers, providing opportunities to vastly extend and improve existing datasets. Summary This study explores the use of large language models (LLMs) to extract structured ethnopharmacological data on medicinal plants and fungi from scientific literature. This will feed into Plants for Health (PfH), an initiative expanding upon Kew's Medicinal Plant Names Services (MPNS). We filtered the CORE dataset of open‐access scientific literature and selected an initial subset of 10 papers. These papers were annotated to highlight scientific plant and fungi names, including the medical conditions they are used to treat and the medicinal effects they are reported to have. Claude 3.5 Sonnet, DeepSeek V3, GPT 4o and Llama 3.1 were evaluated on this data and compared to a baseline SciBERT model. The best‐performing LLM was then fine‐tuned and re‐evaluated, before being deployed on an example subset of 5 more papers and validated using a manual verification tool. The GPT 4o model performed best, achieving an f1 score of 0.63 in the extraction of scientific name‐medical condition/medicinal effect relations, compared to the next best model, DeepSeek V3, with 0.53. When further fine‐tuned and deployed on an example subset of 5 papers, the GPT 4o model achieved a precision of 0.85 and correctly identified 284 scientific name‐medical condition pairs and 363 scientific name‐medicinal effect pairs. Our findings underscore the utility of LLMs in automating the extraction of structured data, offering a scalable approach to augment traditional databases. By improving the breadth of ethnopharmacological information, this work supports better‐informed research and policy on natural products, strengthens the evidence base for traditional medicine and enhances the potential for novel drug discovery.